Do Temporary Help Agencies Help? Temporary Employment Transitions for Low-Skilled Workers

RAQUEL CARRASCO

ISMAEL GÁLVEZ-INIESTA

BELÉN JEREZ

Estudios sobre la Economía Española 2022/11

Mayo de 2022

fedea

Las opiniones recogidas en este documento son las de sus autores y no coinciden necesariamente con las de Fedea.

Raquel Carrasco (UC3M)

Ismael Gálvez-Iniesta (UIB)

Belén Jerez (UC3M)

La industria de servicios privados de empleo generó unos ingresos de unos 495 mil millones de euros en todo el mundo en el año 2019. Dicha industria ofrece una amplia variedad de servicios de intermediación laboral, siendo su actividad principal el llamado “empleo a través de agencias”. Las empresas de trabajo temporal (EET’s) ofrecen a sus empresas clientes la mano de obra de los trabajadores temporales que contratan, mientras son responsables de todas las obligaciones laborales para con los mismos.

El empleo a través de ETT’s ha experimentado un crecimiento sustancial en la mayoría de los países desarrollados desde la década de 1990. El número de trabajadores aumentó de 8,9 millones en 2006 a 49 millones en 2019. El 39% de estos trabajadores se concentra en Europa y nueve países europeos se encuentran entre los quince mercados más grandes del mundo. E crecimiento del empleo a través de ETT’s ha sido particularmente pronunciado en países con mercados de trabajo históricamente regulados, y es en parte consecuencia de la desregulación de esta forma de empleo con el fin de dinamizar los mercados laborales. Se suele argumentar que las ETT’s ofrecen una mayor flexibilidad a las empresas para adaptarse a las fluctuaciones de demanda y reducen los costes laborales (como por ejemplo, prestaciones, pensiones, licencias por enfermedad y vacaciones).

Por todo ello, en las últimas dos décadas se ha generado un intenso debate académico y político sobre cómo este tipo de empleo afecta a las condiciones laborales de los trabajadores. Este artículo presenta un análisis empírico del efecto de tener un contrato a través de una ETT, en lugar de un contrato temporal directo, sobre la duración del empleo temporal. Nuestras estimaciones permiten estudiar el efecto de las ETT’s sobre la probabilidad de transición a diferentes destinos laborales: empleo permanente, empleo temporal, o desempleo. Utilizamos datos administrativos españoles (Muestra Continua de Vidas Laborales) y nuestro periodo de estudio abarca la expansión económica (2004-2007), la crisis económica (2008-2012) y el periodo de recuperación (2013-2017). Así, podemos evaluar el efecto del empleo a través de ETT’s en tres fases diferentes del ciclo económico.

España es un caso particularmente interesante, dadas sus elevadas tasas tanto de desempleo como de empleo temporal. El mercado español de ETT’s ocupó el puesto 12 a nivel mundial en 2019, generando unos ingresos de 5.600 millones de euros. Su participación en el empleo total fue del 4% en 2019, aunque su peso en el número de nuevos contratos firmados fue mucho mayor, alrededor del 10%. Estas cifras son sustancialmente más elevadas entre los trabajadores de baja cualificación. Por ello, nuestro análisis empírico se centra en una muestra relativamente homogénea de trabajadores que tiene como máximo estudios de nivel secundario. Además, nuestro modelo también tiene en cuenta el hecho de que los trabajos temporales en España tienen duraciones cada vez más cortas.

Nuestros resultados indican que tener un contrato temporal a través de una ETT en lugar de un contrato temporal directo aumenta la probabilidad de entrar al desempleo. También encontramos que estos trabajadores tienen más probabilidad de pasar al empleo permanente que los temporales directos, aunque este tipo de transición es muy poco frecuente para ambos. Por otro lado, la probabilidad de que los trabajadores de las ETT’s transiten a otro empleo a través de la ETT es muy alta, mientras que este tipo de transición es casi inexistente entre los temporales directos. A su vez, estos últimos son mucho más propensos a encontrar otro trabajo temporal directo que los trabajadores de ETT’s. Finalmente, encontramos que el efecto positivo de las ETT’s sobre la probabilidad de transición al empleo permanente es procíclico. Por el contrario, el efecto positivo sobre la probabilidad de entrar en el desempleo aumentó durante la recesión en relación con la expansión anterior, pero se ha mantenido elevado durante la recuperación. Un resultado similar se encuentra para las transiciones a otro trabajo a través de ETT, por lo que parte del aumento de dicho tipo de transiciones en los últimos años parece ser más estructural que cíclico.

En resumen, los empleos a través de ETT en España se caracterizan por un mayor riesgo de desempleo y una mayor persistencia que los temporales directos, y estas diferencias se han intensificado en los últimos años. Identificamos principalmente dos tipos de efectos: un pequeño efecto “trampolín” sobre la probabilidad de entrar al empleo permanente, pero también un efecto positivo sobre la probabilidad asociada a tener un nuevo empleo temporal a través de ETT. El primer efecto es consistente con el hecho de que las ETT’s pueden actuar como un mecanismo que mejora la asignación de los trabajadores a los puestos de trabajo. El segundo efecto puede explicarse a través de dos canales: (i) un canal directo, ya que el empleo en ETT está asociado a una elevada probabilidad de transitar a otro empleo a través de ETT, y (ii) un canal indirecto, ya que los trabajadores de ETT’s también tienen una mayor probabilidad de terminar en el desempleo, lo que puede llevarles a reiniciar su búsqueda de empleo a través de una ETT para salir del mismo. Este mayor riesgo de desempleo podría estar relacionado con las diferencias en el tipo de trabajos que los empleadores buscan cubrir cuando contratan trabajadores temporales directamente y cuando lo hacen a través de una ETT, sobre todo si estos últimos son en promedio empleos más inestables caracterizados por tasas de separación más altas. Desafortunadamente, nuestra base de datos no proporciona información sobre los el tipo de trabajos que realizan los trabajadores de las ETT’s. Sin embargo, tenemos evidencia indirecta sobre sus salarios. Los datos indican que, condicional a la duración en el empleo, los empleos a través de ETT tienden a pagar salarios iniciales más elevados. Este hecho es consistente con la idea de que los trabajadores de ETT’s deben ser compensados por la mayor rotación a la que se enfrentan.

Raquel Carrasco^„ Ismael Gálvez-Iniesta Belén Jerez^§

May 8, 2022

Abstract

We investigate how being employed by a Temporary Help Agency (THA) afects transi tion rates to alternative labor market states for low-skilled workers. Our approach is based on the estimation of competing risk discrete duration models, and reveals the importance of accounting for short duration dependence. We use Spanish administrative data for the period 2005-2017. We find that having a THA contract rather than a direct-hire temporary contract increases the probability of entering into unemployment or another agency job at all durations. Agency workers are more likely to transition to permanent employment than their direct-hire counterparts, but these transitions are very infrequent for both. The positive efect of THA employment on the probability of transitioning to a permanent job is procyclical. By contrast, the positive efect on the probability of entering unemployment (or another agency job) increased during the Great Recession relative to the previous economic expansion, and has remained high during the recovery. In words, agency jobs in Spain are characterized by higher unemployment risk and persistence than regular temporary jobs, and these diferences have intensified in recent years. Accouting for unobserved heterogeneity does not alter our main results.

JEL classification: J2, J4, J24, J62, C34.

Keywords: Temporary Help Agency, temporary employment, competing risk duration mod els, unobserved heterogeneity.

^*We thank Ernesto Villanueva, José Ignacio García-Pérez and Maia G¨uell for their comments. We gratefully acknowledge the support from MINECO/FEDER (ECO2015-65408-R, ECO2016-76818, RTI2018-095231-B-I00 and PID2019-107161GB-C31). The usual disclaimer applies.

^„Department of Economics, Universidad Carlos III de Madrid. E-mail: rcarras@eco.uc3m.es

Department of Applied Economics, Universitat de les Illes Balears. E-mail: i.galvez@uib.es

^§Department of Economics, Universidad Carlos III de Madrid. E-mail: mjerez@eco.uc3m.es

1 Introduction

The private employment services industry generated revenues of 495 billion euro worldwide in 2019, according to the international association of private employment services providers (World Employment Confederation, 2021). The industry provides a variety of labor intermediation services, including support services related to the direct recruitment process of its client firms. Yet its core activity is agency work.¹ Temporary help agencies (THAs) supply their clients with temporary workers. That is, they sell the labor of the temporary workers they hire to client firms, while they remain liable for all employment obligations. Agency workers mainly concentrate in the services and manufacturing sectors.

THA employment has experienced substantial growth in several developed countries since the 1990s. The number of agency workers placed worldwide increased from 8.9 million (in full time equivalent) in 2006 to 49 millon in 2019 (World Employment Confederation, 2021). Europe is the largest region in the market for THA services, accounting for a 39% share of worldwide revenues in 2019, and nine European countries rank within the largest fifteen markets worldwide. The growth of THA employment has been particularly steep in countries with historically highly regulated labor markets that have recently been liberalized, and is in part the consequence of deregulation of this form of employment aimed at making labor markets more dynamic (e.g. see Autor, 2009; Fudge and Strauss, 2013; Voss et al., 2013). It is commonly argued that agency work ofers greater flexibility to firms in adjusting to fluctuating demand and reduces non-wage labor costs (e.g. benefits, pensions, sickness and holiday pay entitlements,...), among other things. It is important to note that, while THA employment constitutes a relatively small share of total employment, its share in terms of jobs created (new contracts)² is much higher in most European countries (see Figure 1).

The recent surge of agency work and the fact that the industry has become an increasingly important employer of low-skilled workers over the last two decades has generated an intense academic and policy debate as to how agency jobs afect worker labor market outcomes. It has also been argued that, regardless of its employment share, the THA industry may be shaping the labor market in important ways (e.g. see Fudge and Strauss, 2013). Theodore and Peck (2002)

¹Agency work accounted for 79% of the industry’s total revenue in 2019, while the corresponding share was 13% for direct recruitment services. The industry provides other services associated to the recruitment and management of its client firms’ workforce, but the associated revenue shares are small.

²We indistinctly use the term “job” and “contract” to refer to a contractual relationship between a worker and an employer.

stress that THAs are “active institutional agents in the remodeling of labor market norms and conventions” and play a particularly important role in shaping conditions at the bottom end of the labor market. Harrington and Velluzzi (2008) argue that THA employment contracts shift risk from the firm to the worker. In this study, we find that the use of THAs contributes to the surge of jobs with very short-term duration in Spain.

This paper presents an empirical analysis of the efect of holding a THA contract rather than a direct-hire contract on temporary employment duration. Our approach is based on the estimation of competing risk discrete duration models that allow us to study the efect of THA employment on the probability of transitioning to diferent labor market states: permanent or temporary jobs, or unemployment. We control for personal characteristics and business cycle efects and also consider an extended version of the model allowing for individual unobserved heterogeneity. We use Spanish administrative data (the Continuous Sample of Working Lives). Our sample period covers the economic expansion (2004-2007), the economic crisis (2008-2012) and the recovery period (2013-2017), allowing us to assess the efect of THA employment in three diferent phases of the business cycle.

Spain is a particularly interesting case, given its chronically high rates of unemployment and temporary employment. The Spanish agency work market ranked 12th worldwide in 2019, generating 5.6 billion euros in revenues (World Employment Confederation, 2021). In terms of its share of total employment, it was 4% in 2019, though it accounted for 10% of all new jobs (see Figure 1). These shares are substantially higher among low-skilled workers. For this reason, our empirical analysis focuses on a relatively homogenous sample of male high school dropouts. Due to the prevalence of temporary jobs in Spain with increasingly shorter durations (Felgueroso et al., 2018), our model also accounts for this characteristic of temporary contracts.

We find that having a THA contract rather than a direct-hire temporary contract increases the probability of entering to unemployment at all durations. Our results indicate that agency workers are also more likely to transition to permanent employment than their direct-hire coun terparts, although these transitions are very infrequent for both. On the other hand, agency workers are very likely to transition to another agency job, while this type of transition is almost nonexistent among direct-hire temps. In turn, the latter are more prone to take up another regular temporary job than agency workers. We also find that the positive efect of THA employment on the probability of transitioning to permanent employment is procyclical. By contrast, the positive efect on the probability of entering unemployment increased during

Figure 1: THA employment in Europe, 2019

Share in total employment

Share in total number of contracts

Share in total number of contracts, by skill

Notes: EU-LFS, 2019.

the recession relative to the previous expansion, but it has remained high during the recovery. A similar result holds for transitions to another agency job, so part of the increase in both transitions in recent years seems to be structural rather than cyclical. In summary, agency jobs in Spain are characterized by higher unemployment risk and persistence than regular temporary jobs, and these diferences have intensified in recent years. The estimates of the model with unobserved heterogeneity do not significantly alter our main conclusions.

The rest of the paper is organized as follows. Section 2 provides a summary of the related literature. Section 3 presents some relevant features of the Spanish agency market and Section 4 describes the data. Section 5 formulates the econometric models and estimation procedure. Section 6 discusses the estimation results, and Section 7 concludes.

2 Related Literature

Measuring the efect of THA employment on the individual labor market outcomes of agency workers is challenging. As Autor (2009) emphasizes in a literature survey, “it is inherently dificult to diferentiate the efects of holding given job types from the skills and motivations that cause workers to hold these jobs initially”. It is also dificult to make international comparisons in this context, since the definition of what constitutes agency work difers across countries, and so does the degree of regulation on this type of employment (e.g. see Voss et al., 2013).

While there is widespread evidence that agency work serves as a route out of unemployment (for the less-skilled, in particular), the debate on the THA industry often centers around the question of whether or not agency jobs are a “stepping stone” to more stable forms of employment. Diferent authors use diferent empirical strategies to shed light on this issue, and the resulting empirical evidence is mixed. For instance, Kvasnicka (2009) finds no evidence of a “stepping stone” efect of agency jobs in Germany, and reports that workers entering THA employment from unemployment are more likely to be agency workers over the next four years. Related work by Antoni and Jahn (2009) finds that German agency workers tend to experience repeated spells in agency jobs. According to Ichino et al. (2005), Italian agency workers are less likely to be hired on a permanent basis in the near future than observationally equivalent temporary workers who are directly hired by their employers. By contrast, Zijl et al. (2011) and Jahn and Rosholm (2014) do find evidence of a stepping stone efect in the Netherlands and Denmark, respectively.

Other studies find that spells in THA employment have no significant efects on the sub sequent employment trajectories of agency workers, but do find evidence of selection (e.g. see Autor (2009) and the literature cited therein). In particular, agency workers who are more vulnerable seem to fare worse than their-direct hire counterparts in terms of their prospects of accessing more stable forms of employment. The findings in Autor and Houseman (2005) indicate that this is the case for former welfare recipients in the U.S., for instance. By contrast, using data for Missouri, Heinrich et al. (2005) find that, whereas agency work is an important source of employment for unemployed workers with low skills, it has no measurable efect on the workers’ subsequent employment outcomes, after controlling for unobserved heterogeneity. Yet they also report that workers who experience repeated agency employment spells experience substantial earnings losses in the long run.

Somehow surprisingly, there are only a few studies for Spain. These studies focus on the initial (pre-recession) period after which Spanish THAs were first allowed to operate in the mid 1990s. The analysis in García-Pérez and Muñoz-Bullón (2005a) confirms that THAs facilitate unemployment to work transitions. These authors compare the transitions out and into unem ployment of young workers who had previously held agency jobs to those of similar workers (both in terms of observables and unobservables) who had not. They find that the former workers are more likely to exit unemployment than the latter (at short durations), but they also face a higher risk of transitioning back to unemployment. Moreover, the reported employment hazard rates are higher for workers with low qualifications, and increase with the surge of fixed-term contracts in the late 1990s. Regarding the stepping stone efect, García-Pérez and Muñoz Bullón (2005b), Malo and Muñoz-Bullón (2008) and Bullón and Rodes (2004) find that taking up an agency job instead of staying unemployed may facilitate a future transition to permanent employment for some workers. In particular, controlling for self-selection, García-Pérez and Muñoz-Bullón (2005b) report that this is the case for high-skill workers, though the opposite result holds for workers in low occupational categories, and suggest that this is evidence that Spanish THAs may be playing a role as a screening mechanism that allows firms to select more productive workers. The contribution by Amuedo-Dorantes et al. (2008) difers from the above papers because of its choice of direct-hire temps (instead of unemployed workers) as the control group. These authors find that workers employed through a THA at some point between 1998 and 2004 are less likely to be hired on a permanent basis following their temporary assignment than their direct-hire counterparts, and also report important diferences across groups. We elaborate on the diferences between our work and that of Amuedo-Dorantes et al. (2008) and García-Pérez and Muñoz-Bullón (2005b) in detail in Section 7, after we discuss our main results.

3 Temporary Help Agencies in Spain

3.1 Institutional and legal framework

The Spanish labor market stands out not only for its chronically high unemployment rate (specially among the youth) but for having one of the highest shares of temporary employment in Europe as well. At the onset of the Great Recession, about one third of all salaried employees in Spain were temporary. This share has been relatively stable since the late 1990s.

The high incidence of temporary employment in Spain has been attributed to the liberal ization of the use of fixed-term contracts in 1984 (see Bentolila et al., 2012; Dolado et al., 2002;

García-Pérez et al., 2019). It has been argued that the regulation of these contracts in Spain is less stringent than in other European countries with similar labor market institutions and its de-facto enforcement is limited in reality, whereas the regulation of permanent contracts is much more strict (e.g. Bentolila et al., 2012). The share of temporary employment has remained high in recent years, despite the introduction of latter countervailing reforms which failed to reduce labor market dualism. As noted by Felgueroso et al. (2017), the deepest reform, which took place in 2012, significantly reduced the degree of employment protection of permanent workers, which is now close to the OECD average. Still, almost 85% of all contracts signed in 2017 were temporary. Felgueroso et al. (2017) also highlight the fact that these contracts are of increasingly shorter durations.

Unlike regular temporary employment, agency work had traditionally been tightly regulated in Spain. THAs were first authorized and regulated by the Spanish Parliamentary in 1994 (Law 14/1994). According to the statutory definition, their “activity is intended to transfer hired employees to user companies on a temporary basis”. The law specifies a triangular employment relationship, whereby an agency and a worker sign an employment contract (“contrato de trabajo”), and the agency and the user company sign a contract specifying the terms under which the worker is transferred to the user company (“contrato de puesta a disposición”).³ Once the worker joins the user company, the latter assumes the direction and control of employment activity. While the assignment to the user company is temporary, employment contracts signed with THAs can be either temporary or permanent, though the latter are rare in practice (see Countouris et al., 2016).

The main objective of the 1994 law was to standardize the regulation of agency work with that already existing in some European countries (along the lines of the 96th ILO Convention), and at the same time to safeguard the labor rights and employment protection of agency work ers. A number of requirements were established for the operation of THAs, such as special administrative permissions, a minimum organizational structure, the fulfillment of salary and social security obligations, and exclusive dedication to temporary agency work activities. The use of agency work was also banned in certain situations,⁴ and so were training agency contracts.

³These contracts must include detailed information about the reasons for the assignment and the job to be performed, remuneration and estimated duration of the assignment, job location and time schedule, and collective bargaining agreements applying to both the user company and the agency. THA collective agreements in Spain are intersectoral and national, and compatible with other agreements at regional level.

⁴Specifically, the hiring of agency workers to replace workers on a legal strike, carry out particularly hazardous jobs, fill vacancies caused by recent layofs at the user company, or transfer workers to another agency was rendered illegal.

The law requires that THAs provide adequate training for their employees before they join the user company and devote at least 1 percent of their total wage bill to training, and forbids employees to bear any training costs. THA employment contracts cannot include clauses pro hibiting the hiring of a worker by the user company once the contract expires. Agency workers who continue to work at the user company after their contract expires are automatically con sidered permanent workers of the latter (just as the users company’s own temporary workers). The law also specifies other basic worker rights (e.g. regarding security and health at work) that user companies must comply with.

In 1999, the law was amended to ensure the principle of equal pay of agency workers and to strengthen their employment protection (Law 29/1999). In particular, the amendment ruled that agency workers ought to receive at least the wage rate associated with their position as specified by the collective bargaining agreement of the user company.⁵ At the same time, it contributed to the deregulation of this sector by allowing THA employment contracts to be signed “in the same cases and under the same conditions and requirements” which apply to regular fixed-term contracts, and by weakening some of the original restrictions in the 1994 law.

In the aftermath of the 2008 financial crisis, and in the context of the progressive deregulation that aimed at enhancing labor market flexibility, additional restrictions on the use of agency work were lifted. In particular, the ban on sectors such as construction and mining, as wel as in public administration was removed in 2010, formally transposing EU Directive 2008/104 (Law 35/2010). The current THA collective agreement contains a commitment of unions to avoid the inclusion of clauses in any collective agreements that exclude the use of temporary help services.⁶ THAs were also allowed to act as placement agencies and sign training contracts in 2014, and the authorization process was greatly simplified (Act 18/2014).

3.2 THA employment in Spain: descriptive evidence

We use data from the Continuous Sample of Working Lives (Muestra Continua de Vidas Lab orales, MCVL hereafter), which is a micro-level dataset extracted from Spanish Social Security records. An exhaustive description of the dataset is provided in Section 4.

Figure 2 displays the evolution of THA employment during the period 2005-2017 for workers aged 16 to 65 (solid line), and for a restricted sample composed by male high school dropouts aged 20 to 45 (dashed line). The empirical analysis in Section 5 is based on the restricted sample because, as we discuss below, the incidence of THA employment is substantially higher among these workers. For both samples, we report the share of THA employment in salaried employment, and also the respective share of THA contracts in the total number of contracts signed and terminated in a given month. While the first statistic provides a static measure of the importance of agency work in total employment, the other two speak to its importance for job creation and job destruction flows and thus for the dynamics of the Spanish labor market.

⁵As opposed to user company employees, agency workers do not receive holiday periods or extraordinary payments, but the proportional payments are included in their wages

⁶Before 2011, many collective agreements included restrictions on the use of agency work. Several of these restrictions were revoked by Law 35/2010. See Countouris et al. (2016).

In 2017, THA employment accounted for around 1% of salaried employment in the full sample and for 3% in the restricted one (upper panel of Figure 2). The upward trend in these shares indicates that this type of employment has gained importance in recent years. In the restricted sample, the share fell sharply during the early phase of the Great Recession but increased substantially afterwards, suggesting that agency work has become more relevant in the low-skill segment of the labor market during the long recession (when unemployment rate was high and workers struggled to find jobs). Still, these moderate employment shares conceal the actual importance of agency work for job creation and job destruction. Around 10% of all contracts signed in a given month in 2017 were agency contracts (bottom left-hand panel of Figure 2). This share doubles in the restricted sample, where one out of five contracts were signed with an agency. Regarding job destruction, 12% of all contracts terminated in a given month in 2017 where agency contracts, the corresponding share shooting up to around 22% in the restricted sample (bottom right-hand panel). Again, all the aforementioned shares display an increasing pattern since 2009.

Table 4 in the Appendix displays the composition of workers by the type of contract, for 2017. The age distribution of temporary workers in agency and direct-hire contracts is similar, but these workers are significantly younger than those on permanent contracts. On the other hand, the share of males in agency contracts (59%) exceeds the corresponding shares in directhire temporary and permanent contracts (around 50%). This is consistent with the fact that part-time is less common in agency contracts (25%) than in direct-hire temporary jobs (29%). The share of foreign workers in agency contracts (16.2%) is slightly higher than in direct-hire jobs (13.1%) and almost identical than that in permanent contracts (16.3%). Low-educated workers are also over-represented in agency jobs. In particular, high school dropouts account for 64% of all agency contracts, while the corresponding share is 57.3% for direct-hire temporary contracts and 55.4% for the permanent ones.

Figure 2: Evolution of THA employment THA share in salaried employment THA share in new contracts THA share in terminating contracts

Notes: THA contract shares defined as the % of THA contracts over all contracts signed/terminated in a given month. Shaded areas indicate recessions (Double-Recession: 2008Q2-2009Q2, 2010Q4-2013Q2). All series are seasonally adjusted. Source: MCVL, 2005-2017.

There are also important diferences in the composition of agency and regular temporary jobs by skill type and job duration. Strikingly, 96% of all agency contracts correspond to medium-low and low-skilled occupations.⁷ These occupations are not nearly as relevant in the case of direct-hire temporary and permanent contracts (73% and 64%, respectively). In fact, whereas agency contracts account on average for 12% of all contracts signed in 2017, this share is close to 20% for jobs in medium-low and low-skilled occupations.⁸ The other key diference is that agency contracts are shorter than regular temporary contracts (Figure 3). For instance, 33% of all THA contracts signed in 2017 lasted only 1 day and 82% lasted less than 1 month, whereas the respective shares are 16% and 53% for regular temporary contracts. Therefore, the THA industry seems to be important in explaining the surge of jobs with very short-term durations (e.g. it accounted for 23% of all 1-day jobs and for roughly 20% of jobs lasting less than a month in 2017, as shown in Table 5).

⁷See Table 7 in the Appendix for the description of the skill categories.

⁸Table 5 in the Appendix reports the shares of THA contracts by worker/job characteristics.

Figure 3: Duration of THA vs direct-hire temporary jobs

Notes: Full sample. Source: MCVL, 2005-2017.

One important drawback of the MCVL is that it does not report the sector of activity of agency jobs. The sectoral composition of the THA workforce, however, can be tracked using the THA statistics provided by the Spanish Social Security (see Figure 7 in the Appendix).⁹ THA contracts mainly concentrate on manufacturing (28%), agriculture, forestry, and fishing (17%), accommodation and food services (14%), transportation and storage (14%), and wholesale (11%). These are sectors where temporary workers abound. Agency contracts are highly over-represented in the manufacturing sector, where they accounted for more than 50% of all contracts in 2017, as well as in the transportation and storage sector, where the corresponding share exceeds 40%.

Summing up, in Spain, THAs are mainly employing young males with low levels of education in low-end occupations. This is the reason why our empirical analysis is based on the restricted sample of male high school dropouts aged 20 to 45.

We close this section by describing the aggregate patterns of employment transitions for agency, direct-hire temporary and permanent contracts within the restricted sample (Figure 4). Let A, D and P stand for agency, direct-hire temporary and permanent employment, respectively, and let U denote unemployment. We compute the proportion of A, D, and P contracts that transition to states A, D, P, and U from month t to month t + 1 (in panels a, b, c and d, respectively). For instance, the line DP (AP ) in Panel a depicts the proportion of direct-hire temporary (agency) contracts that transition to a permanent contract in a given month. The transitions P P , DD, and AA are defined as the proportion of contracts that transition to another s contract. Notice that transitions are defined as a change of contract, and they can be signed with the same employer or with a diferent one.¹⁰

⁹To guarantee comparability of the two data sources, we have computed all the previous descriptive statistics (composition by age, education, occupation, duration...) using data from the THA statistics and compared them to the ones we obtained using the MCVL. The numbers are very similar.

Throughout the whole period, the proportion of exits to U is higher for agency contracts than for direct-hire temporary contracts (Panel b). The opposite is true for the exits to P (Panel a). Yet transitions into (and out of) permanent contracts are rare events since few workers hold permanent jobs in our sample. Note that the largest gap is observed in the proportion of exits to A: throughout the whole period this proportion was above 30% in agency contracts and around 1% in direct-hire temporary contracts (Panel d). The opposite pattern is observed in the transition to D, though the gap is smaller; e.g. the proportion is around 15-18% for regular temporary contracts and about half that number for agency contracts during the most recent period (Panel c). In any case, all these figures must be taken with caution since, as we have shown, agency and direct-hire temporary workers have diferent characteristics and agency employment spells also tend to be shorter.

4 Data

4.1 Data and sample descriptive statistics

The MCVL is a 4% representative random sample of all workers afiliated with the social security administration (working as employed or self-employed, receiving a public pension or being registered as unemployed) during at least one day in the year when the sample is extracted. The dataset was released in 2004 and, after that year, it follows the same sample of individuals over time, adding new observations each year to replace absences while keeping the sample representative of the population. Furthermore, the data provides retroactive information on the workers’ entire labor market history, allowing us to construct accurate measures of experience.

We pool the database for years 2005-2017, so our dataset keeps track of the complete em ployment history of any worker who had registered at least one day of activity in the socia security administration between 2005 and 2017.¹¹ Additional information is available for each individual, including personal and demographic characteristics (age, gender, education, nationality, region of residence), employer information at the establishment level (location, size), and labor market information (industry, occupation, type of contract). The dataset also reports whether the worker is employed by an agency. Yet no information regarding the user companies (and thus the industry) where agency workers perform their work is reported. The unit of obser vation is the spell in the individual’s labor market history defined as a contractual relationship between a firm and a worker, reporting the starting and ending date of the spell.

¹⁰The percentage of transitions to another temporary contract with the same employer is 87.3% for AA and 59.26% for DD.

Panel a. To permanent job Panel c. To direct-hire temp job

Panel b. To unemployment Panel d. To THA job

Notes: Sample of low-educated male workers aged 20-44. Source: MCVL, 2005-2017.

Since our sample is only representative from 2005 on, we restrict to employment spells that started in 2005 or later. Given that we are interested in temporary employment, we exclude self-employed and permanent spells (contracts). We also restrict to contracts with the general social security regime, abstracting from agricultural and other exceptional regimes. The restricted sample of male high school drop-outs aged 20 to 44 consists of 168,897 individuals. To save computational memory, for our estimations we draw a random subsample of 5,000 individuals, which leaves us with 44,393 contracts (495,347 observations). Out of these, 7,859 (17.70%) correspond to agency contracts, whose characteristics are reported in Table 1. While the workers’ composition of agency contracts and direct-hire temporary contracts is similar, some of the key diferences described in Section 3.2 prevail within this more homogeneous sample. Agency workers are slightly younger and more concentrated in low-skilled occupations, and they experience shorter employment spells than their direct-hire counterparts (3.95 vs 10.77 weeks, on average).

¹¹Data extraction for each year is done in April of the following year. Therefore, we have some spells in our dataset ending between January 2018 and April 2018.

Figure 5 allows us to take a first glimpse at the diferences between THA employment (solid line) and regular temporary employment (dashed line) in terms of the likelihood of leaving current temporary employment by duration. In particular, for each contract duration measured in weeks, we calculate the ratio of the number of agency and direct-hire temporary contracts that exit to states P , U , D and A to the total number of contracts of each type that are still active.

Note that all figures display peaks around the time of contract termination (the most promi nent peak being typically associated with 6-month spells).¹² As one would expect, and except for the peaks, the exit rate to unemployment falls sharply with duration for both agency and regular temporary contracts during the first 2 months (Panel b of Figure 5). By contrast, there is no duration dependence in exits to permanent employment (Panel a of Figure 5).¹³ As to the diferences by contract type, average exit rates to states U and P are both higher for agency contracts than for regular temporary contracts. Regarding transitions within temporary employment, direct-hire temps are more likely to switch to another regular temporary contract than to an agency contract (Panel c of Figure 5), while the opposite is true for agency workers (Panel d). The diferences in the latter transition are large: the average exit rate to state A is

¹²The maximum legal duration of fixed-term contracts for circumstances of production (“contratos eventuales por circunstancias de la producción”), which account for about 44% of the spells in the sample, is 6 months (within a 12-month period). While collective wage agreements may modify this maximum limit, it cannot exceed 12 months (within any 18 month period). The maximum duration of per-task fixed-term contracts (“contratos de obra y servicio”), which represent about 50.2% of the spells in the sample, is 3 years.

¹³Carrasco and García-Pérez (2015), among others, highlight the negative duration dependence in exit rates to unemployment for temporary contracts in Spain, and show that there is no duration dependence in these exit rates for permanent contracts.

Table 1: Descriptive statistics

Source: MCLV, 2005-2017.

15.7% for agency workers, but it is less than 0.1% for direct-hire temps.

The average hazards shown in Figure 5 result from the aggregation of individuals with diferent hazard rates in diferent time periods. With the estimation of the econometric model, controlling for personal characteristics and time efects, we should be able to disentangle from the observed hazards the efects that result from individual heterogeneity.

Figure 5: Kaplan hazard rates by job duration Panel a: To permanent job Panel c: To direct-hire temp job

Panel b: To unemployment

Panel d: To THA job

Notes: Sample of low-educated male workers, aged 20-44. Source: MCLV, 2005-2017

5 Empirical models

We analyze the dependence of the exit from temporary employment on the type of contract (direct or through an agency), the length of time employed and on other economic and demographic variables by the estimation of duration models. We treat temporary employment duration (T ) as a discrete random variable which is right censored when the individual is still employed at the time of leaving the sample.

Note that our empirical approach does not allow us to infer causality. Our estimates result from the interplay of labor demand and supply factors, but they are useful to have a deeper knowledge of the characteristics associated with the diferent temporary employment risks that agency workers and direct-hire temps face.

5.1 Single risk models

We consider first a single-risk discrete duration model. Formally, let be the length of a spell of temporary employment. At this stage, only one hazard exists which may cause failure: exiting a temporary employment spell, without distinguishing between destination states. The discrete hazard rate at t is defined as

\[h _ {i} (t) = \operatorname * {P r} \left[ T _ {i} = t \mid T _ {i} \geq t, T H A _ {i} (t), x _ {i} (t) \right] = F (\alpha_ {0 t} + T H A _ {i t} \alpha_ {1} + x _ {i t} ^ {\prime} \alpha_ {2}).\tag{1}\]

In this expression, is the logistic cumulative probability function, , α_0t is a specific parameter associated to each duration t that captures additive duration dependence in a flexible way, and T HA is a dummy variable taking the value 1 if the individual is employed through an agency and 0 if she holds a regular temporary contract. Additional efects of duration are captured by introducing the interaction of T HA with duration as a regressor.¹⁴ We condition on a vector of variables which includes age, immigrant status, a dummy for part-time employment, occupational skill level, contract type, previous experience in permanent, agency and direct-hire temporary employment, and year (see Table 6 in the Appendix for the definition of the variables).

If is the observed duration of the spell (completed or censored) and is an indicator variable equal to 1 if the spell is completed and 0 if it is censored, its contribution to the log likelihood function is given by

\[\begin{array}{c} \log L _ {i} = \left[ c _ {i} \left\{\sum_ {t = 1} ^ {d _ {i} - 1} \log [ 1 - h _ {i} (t) ] + \log h _ {i} (d _ {i}) \right\} + \right. \\ \left. (1 - c _ {i}) \left\{\sum_ {t = 1} ^ {d _ {i}} \log [ 1 - h _ {i} (t) ] \right\} \right]. \end{array}\tag{2}\]

The log-likelihood function of the sample is given by the sum of the contributions of the N temporary employment spells and is maximized with respect to the model parameters to provide Maximum Likelihood (ML) estimates.

It is well-known that this model can be estimated as a sequence of binary choice equations with cross-equations restrictions defined on the surviving population at each duration t (e.g. see Bover et al., 2002; Jenkins, 1995; Narendranathan and Stewart, 1993). More specifically,

¹⁴Notice that in this type of non-linear models the marginal efects of the variables are non-constant even if one does not include interactions between the variables in the model specification. We include this interaction just to capture potential additional efects of THA by duration.

for each t we define the following variables:

\[\begin{array}{l} y _ {i t} = 1 (T _ {i} = t), \\ S _ {i t} = 1 (T _ {i} \geq t), \end{array}\]

where takes the value 1 if an exit from the temporary employment spell is observed in period t and 0 if there is no exit or if the observation is censored at t, and takes the value 1 for the surviving population at each t (notice that with probability one). We can then write

\[h _ {i} (t) = \operatorname * {P r} \left[ y _ {i t} = 1 \mid S _ {i t} = 1, T H A _ {i} (t), x _ {i} (t) \right].\]

Therefore, the log-likelihood function can be written as:

\[\begin{array}{c} \log L (\alpha) = \sum_ {i = 1} ^ {N} \sum_ {t = 1} ^ {d _ {i}} S _ {i t} \left\{y _ {i t} \log F (\alpha_ {0 t} + T H A _ {i t} \alpha_ {1} + x _ {i t} ^ {'} \alpha_ {2}) \right. \\ \left. + (1 - y _ {i t}) \log (1 - F (\alpha_ {0 t} + T H A _ {i t} \alpha_ {1} + x _ {i t} ^ {'} \alpha_ {2})) \right\}, \end{array}\tag{3}\]

where each spell i contributes observations. Since the hazard rate in the likelihood function is a logistic distribution, estimation is equivalent to estimating a sequence of logit models (with cross-equation restrictions) defined on the surviving population at each duration.

5.2 Competing risk models

The previous model considers the determinants of a single risk, that of exiting a temporary employment spell. But much of the interest comes from the analysis of a situation where there are competing risks. Therefore, we consider a model in which there is more than one possible exit from a temporary employment spell and allow for a diferent behavior with respect to the diferent risks. Specifically, we distinguish between exits to unemployment (U), permanent employment (P ), agency employment (A) and regular temporary employment (D). As already noted, exits to occur when workers employed in a regular temporary contract (agency contract) sign another contract of this type with the same firm (agency) or with a diferent one.

We define exit rates to each of the states conditional on not exiting to the alternatives state:

\[\begin{array}{r c l} h _ {i j} (t) & = & \operatorname * {P r} \left[ T _ {i} = t, j _ {i} = 1 \mid T _ {i} \geq t, s _ {i} = 0 \right], \\ j & = & U, P, A, D; \text {for all} s \neq j, \end{array}\tag{4}\]

where are the alternative’s indicators which equal 1 if an exit into the corre sponding state is observed and 0 otherwise.

We can treat the discrete duration competing risk model as a sequence of discrete choice models by defining the following indicators:

\[y _ {i t j} = 1 (T _ {i} = t, j _ {i} = 1), j = U, P, A, D.\tag{5}\]

Using this notation, we specify the conditional exit rates to each state as binary choice logit models:

\[\begin{array}{r c l} h _ {i j} (t) & = & \operatorname * {P r} \left[ y _ {i t j} = 1 \mid T _ {i} \geq t, T H A _ {i} (t), x _ {i} (t), y _ {i t s} = 0 \right] = \\ & = & \frac {\exp (\alpha_ {0 t} ^ {j} + T H A _ {i t} \alpha_ {1} ^ {j} + x _ {i t} ^ {'} \alpha_ {2} ^ {j})}{1 + \exp (\alpha_ {0 t} ^ {j} + T H A _ {i t} \alpha_ {1} ^ {j} + x _ {i t} ^ {'} \alpha_ {2} ^ {j})}, \\ j, s & = & U, P, A, D; \text {for all} s \neq j. \end{array}\tag{6}\]

We define indicators if spell i ends in state otherwise.¹⁵ The conditional loglikelihood function for exit j is then given by:

\[\begin{array}{r c l} \log L _ {j} & = & \sum_ {i = 1} ^ {N} \left[ c _ {i j} \left(\sum_ {t = 1} ^ {d _ {i} - 1} \log \left[ 1 - h _ {j, i} (t) \right] + \log h _ {j, i} (d _ {i})\right) \right. \\ & & \left. + (1 - c _ {i j}) \sum_ {t = 1} ^ {d _ {i}} \log \left[ 1 - h _ {j, i} (t) \right] \right]. \end{array}\tag{7}\]

The full log-likelihood is the sum of terms like (7) over . Therefore, exits to alternative s in (7) are treated as censored observations, so that formally log has the same form as the likelihood with a single risk of the previous section. This implies that the Conditional ML estimators , defined as the maximum of log , log , log and log , can be obtained as separate ML estimates of four binary logit models.

Once we have obtained the estimates of the coeficients , we estimate the parameters of interest in this type of models, which are the average predicted probabilities and the average marginal efects of the main variable of interest, T HA, on the probabilities of exiting to a specific alternative for each duration.

Bover and Gómez (2004) point out that another possibility is to model the so-called “transition intensities” for each state, instead of the conditional hazard rates, and to measure the efect of the explanatory variables with respect to the transition intensities. In particular, the transition intensity to state is defined as the probability of exiting to that state at among those who remain in that state for at least periods:

¹⁵Note that

\[\phi_ {i j} (t) = \operatorname * {P r} \left[ T _ {i} = t, j = 1 \mid T _ {i} \geq t \right], j = U, P, A, D.\tag{8}\]

The relationship with the conditional hazards is given by:

\[{h _ {i j} (t)} = {\frac {\phi_ {i j} (t)}{1 - \sum_ {s \neq j} \phi_ {i s} (t)}.}\tag{9}\]

If the specification used for the transition intensities is a multinomial logit model as

\[\phi_ {i j} (t) = \frac {\exp (\alpha_ {0 t} ^ {j} + T H A _ {i t} \alpha_ {1} ^ {j} + x _ {i t} ^ {'} \alpha_ {2} ^ {j})}{1 + \sum_ {s} \exp (\alpha_ {0 t} ^ {s} + T H A _ {i t} \alpha_ {1} ^ {s} + x _ {i t} ^ {'} \alpha_ {2} ^ {s})},\]

in accordance with the relationships in (9), the conditional exit rates are binary logit with the same parameters as specified in (6). The estimation of the parameters in this case would require the joint log-likelihood function of a multinomial logit model.

Both methods, modeling the conditional hazard or the transition intensities, provide consis tent and asymptotically normal estimates of the parameters, although the latter is in general asymptotically more eficient. We have chosen to model the conditional hazard rates because the separate estimation of is computational faster, especially when one accounts for the efect of time invariant unobserved heterogeneity among individuals, as we will show in next section. Moreover, given that the sample used in this paper is very large, we do not expect important relative ineficiencies from the separate estimation of .

5.3 Competing risk models with unobserved heterogeneity

In the specification of the hazards (6) there may be omitted variables which could introduce a bias in the estimators of the model parameters. Such unobserved heterogeneity can be related to the job search efort, job networking, motivation, or individual ability that may afect the transition rate out of a temporary employment spell and is likely to be correlated with the type of employment contract and introduce spurious duration dependence.

A version of the competing risk model allowing for unobserved heterogeneity for a given exit state, , would be given by

\[h _ {i j} (t, \eta^ {j}) = \frac {\exp (\alpha_ {0 t} ^ {j} + T H A _ {i t} \alpha_ {1} ^ {j} + x _ {i t} ^ {'} \alpha_ {2} ^ {j} + \eta_ {i} ^ {j})}{1 + \exp (\alpha_ {0 t} ^ {j} + T H A _ {i t} \alpha_ {1} ^ {j} + x _ {i t} ^ {'} \alpha_ {2} ^ {j} + \eta_ {i} ^ {j})}.\tag{10}\]

The problem of how to account for the distribution of , has been addressed extensively in the literature (e.g. see Van den Berg, 2001). Heckman and Singer (1984) proposed to adopt a semi-parametric approach to identify the unobserved distribution from a mixed distribution assuming that is a discrete random variable with finite support. We follow this approach and we assume that has two mass points for a given with associated probabilities and . Therefore, there are two types of individuals whose unobserved characteristics afect each transition diferently: individuals with high and with low exit rates to , and . The probabilities associated to each type also difer according to the transition being considered.

The log-likelihood function with unobserved heterogeneity for exits into state j takes the form:

\[\log L _ {j} (\alpha^ {j}, \eta^ {j}) = \sum_ {i = 1} ^ {N} \sum_ {l = 1} ^ {2} \log L _ {i j} (\alpha^ {j}, s _ {l} ^ {j}) \times P _ {l} ^ {j},\tag{11}\]

where log is given by

\[\begin{array}{l l l} \log L _ {i j} (\alpha^ {j}, s _ {l} ^ {j}) & = & \left[ c _ {i j} \left(\sum_ {t = 1} ^ {d _ {i} - 1} \log \left[ 1 - h _ {i j} (t, s _ {l} ^ {j}) \right] + \log h _ {i j} (d _ {i}, s _ {l} ^ {j})\right) \right. \\ & & \left. + (1 - c _ {i j}) \sum_ {t = 1} ^ {d _ {i}} \log \left[ 1 - h _ {i j} (t, s _ {l} ^ {j}) \right] \right]. \end{array}\tag{12}\]

The mass points are estimated as constant terms in the hazard index, and the associated probabilities of each mass point are also estimated jointly with the rest of the model parameters. For the estimation we use the “gllamm” command in STATA. To obtain the predicted probabilities and the marginal efects of interest, we have computed the corresponding magnitudes for each point of support of the heterogeneity distribution. These magnitudes are then weighted using the estimated probabilities for each point.

6 Estimation results

We report the estimates of the hazard of exiting a temporary employment spell focusing on the efect of the type of contract. We first present the results for each the four risks considered , for each duration and for diferent time periods. A discussion of the results allowing for unobserved heterogeneity then follows. The qualitative impact of the type of contract is discussed in terms of the sign and statistical significance of the estimated coeficients. The size of these impacts is assessed by the estimation of the average predicted probabilities and marginal efects at diferent durations and time periods. While we do not claim causality in our estimations, for the sake of clarity, we will refer to the latter as the marginal efects of THA employment.

Table 8 in the Appendix presents the ML estimates for the single risk model (column 1) and the competing risk model for the four risks considered (columns 2 to 5), while Table 9 reports the estimates of the competing risk model which accounts for unobserved heterogeneity. As already mentioned, we control for duration dependence in a flexible way by including a dummy variable for each weekly duration.¹⁶ Due to the relatively small number of observations, durations longer than 53 weeks are treated as right-censored at 53 weeks.

6.1 Predicted employment hazards by contract type

Column 1 in Table 8 contains the parameter estimates for the single risk model. Our results indicate that employment exits are more likely for agency workers than for their direct-hire counterparts (notice the positive coeficient on the dummy variable T HA). Also, there is a negative duration dependence, that is, the longer an individual has been employed, the more likely she is to continue in this state (see Figure 9 in the Appendix). Nonetheless, this model does not specify separately the efect of the covariates through their efect on the probability of exiting to diferent alternative states. For this reason, we focus on the estimates of the competing risk model (columns 2 to 5).

Our results point to a positive efect of THA employment on the hazard of leaving temporary employment, both to P (column 2) and U (column 3). This positive efect increases with duration for exits to P (notice the positive coeficient on the interaction between the dummy variable T HA and duration), while this interaction is not statistically significant for exits to unemployment. To assess the magnitude of the efect, Figure 6 plots the average predicted hazards for states P (Panel a)¹⁷ and U (Panel b) for both agency and regular temporary contracts.

In the case of state P , the predicted hazard is higher for agency contracts than for regular temporary contracts at all durations. For instance, in the first of employment the respective hazards are 0.51% and 0.23%, and they are 0.50% and 0.20% in the 10th week. Diferences are statistically significant only for job durations shorter than 15 weeks. The estimated average marginal efect (across durations) is around 0.56 pp. This marginal efect is small because, as we have already noted, exits to P are rare in the sample (where the average exit rate is 0.26%). Still, agency workers are three times more likely to exit to P than their direct-hire counterparts, the average estimated probabilities being 0.8% and 0.24%, respectively. Note also that we do not find evidence of duration dependence: as shown in the graph, the hazard levels of, except for the peaks.

¹⁶We do not report the duration coeficients in the table, but all estimations include 53 weekly job duration dummies as additional regressors.

¹⁷For this transition, the estimates are noisier for longer employment spells because there are few observations in the corresponding cells. Consequently, we focus our discussion on short durations.

Panel c: To direct-hire temp job

Panel d: To THA job

Notes: The figures display the average predicted probabilities at each duration. Apart from controlling by agency employment and the interaction between agency employment and duration, all models include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, direct-hire temporary employment, and in agency employment and year fixed efects. The dashes lines correspond to 95 percent confidence intervals. Source: MCVL, 2005-2017.

As to the transitions to U, notice that negative duration dependence is much more evident. In the first week of employment, the estimated probabilities are 10.3% and 8.6% for agency workers and direct-hire temps, respectively, and the implied marginal efect of THA employment is 1.7 pp. This marginal efect decreases with job duration as both probabilities decrease; e.g. it is 0.9 pp for jobs lasting one month. The average marginal efect is 0.40 pp. These results suggest that agency jobs are characterized by higher turnover, especially at short durations, being more unstable than regular temporary jobs.

Consider now the estimated exit probabilities to states A and D (columns 4 and 5 in Table 8). Whereas it is common for agency workers to sign another agency contract after their contracts expire, direct-hire temps seldom do so (see Panel d in Figure 6). Also, the estimated probability for agency workers displays negative duration dependence. This probability is 30.6% within the first week of employment and 10% after 1 month, and it remains above 2% even after 6 months, suggesting that THA employment is particularly persistent. For direct-hire temps, however, the probability is close to zero regardless of job duration. As to transitions to state D (see Panel c in Figure 6), these are more likely for direct-hire temps than for agency workers, especially for durations shorter than 20-25 weeks. For instance, in the first week of employment, the estimated probability is 15.6% for the former and 4.4% for the latter. In general, the implied negative marginal efect of THA employment is larger for short employment spells, amounting to -4.8 pp after 1 month of employment, and it is -0.57 on average.

Efect of contract type over time It is also important to understand how the aforementioned marginal efects change over time. To do so, we split the sample in three subperiods: the economic expansion (2004-2007), the economic crisis (2008-2012) and the recovery period (2013-2017). Table 2 reports the average marginal efects for each subperiod (see also Figure 8 in the Appendix).

The marginal efect associated to employment exits to state P is clearly procyclical: it fell from the expansion to the recession period, rising again in during the recovery. So, while we have already noted that agency workers faced on average a higher (if small) chance of landing a permanent job than direct-hire temps, the diference was higher in the expansion and recovery periods than in the recession. For instance, when employment duration is 3 months, the probability of entering P is 0.5 pp higher for agency workers than for workers on agency contracts during the expansion and recovery, whereas this number falls to 0.3 pp during the recession period.

As far as exits to U are concerned, note that the marginal efect of THA employment increased after 2008, but there are not significant diferences between the recession and recovery periods in this case. In other words, the diferences in unemployment risk between agency workers and direct-hire temps increased at the onset of the recession and have not narrowed during the recovery. For instance, after 1 week of employment, holding an T HA contract (rather than on a regular fixed-term contract) increased the probability of becoming unemployed by 1.3 pp on average before 2008, while the corresponding estimate after 2008 is 1.9 pp.

Interestingly, the positive marginal efect of THA employment on the probability of transitioning to another agency job (A) displays a clear increasing pattern over time, as the agency sector experienced a gradual liberalization. The efect is also larger at short durations in all subperiods, indicating a close relationship between THA employment and short-duration jobs. Note that, for durations of at most 1 week, the marginal efect increased sharply from 21.9 during the pre-crisis period to 34.7 pp in the recovery period. On the other hand, the negative marginal efect of THA employment on the probability of transitioning to D is also substantially larger (in absolute value) for the latest period of the sample (2013-2017). These results suggest that the degree of persistence in THA employment has increased during this period.

Table 2: Marginal efect of agency contract by period

Notes: The table displays the average marginal efects of agency employment by period for a selected set of job duration. Letters E, C, and R stands for the expansion (2004-2007), crisis (2008-2012) and recovery periods (2013-2017), respectively. For durations that are aggregated, we compute the average. Apart from controlling by agency employment and the interaction between agency employment and duration, all estimations include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, in direct hire temporary employment and in agency employment and year fixed efects. Source: MCVL, 2005-2017.

Other characteristics The estimated efect of other variables on the above hazards can be summarized as follows. We find that foreign workers and younger workers (aged 20 to 30) have a lower probability of exiting temporary employment to U, D and A than natives and older workers (aged 30 to 40). There is also evidence of a positive efect of part-time employment and higher job skills on the temporary employment hazard rate to U, D and A. Finally, previous experience in (both agency and regular) temporary employment is associated with a lower probability of exiting temporary employment, while previous experience in permanent employment increases the probability of exiting temporary employment only to P . Notice that, in general, the efects of the covariates on the exit to P are estimated with very low precision given the small number of transitions observed.

Regarding the marginal efect of THA employment by worker characteristics,¹⁸ we do not find significant diferences by age and immigrant status. We find larger marginal efects of THA employment on the exit rates to A and D for low-skilled workers, and no significant diferences for exits to P and U .

6.2 Unobserved heterogeneity

We turn to the estimation results of the competing risk model with unobserved heterogeneity. Before discussing the estimates of the model coeficients, we present the estimates of the parameters related to the unobserved heterogeneity distribution. As previously explained, the distribution is specified with two mass points for each risk. As shown in Table 3, the probability of the type with a high exit rate to P (that is, the probability associated to the smallest mass point in absolute value) is the lowest in our sample, 5.1%. In other words, having the unobserved characteristics that make this transition more likely is rare. A similar result is obtained for exits to A and D. However, the probability of the type with a high exit rate to U is approximately 38%. That is, being an individual with unobserved characteristics that makes this transition more probable is not so unlikely.

The estimates of the log-likelihood function for the hazard model with unobserved heterogeneity are presented in Table 9 in the Appendix. The coeficients of the dummy for the type of temporary contract have the same sign in the models with and without unobserved heterogeneity. Figure 10 in the Appendix plots the average predicted hazards for each of the four destination states for both agency and regular temporary contracts. These have been calculated as the predicted hazards for each individual type weighted by the estimated probability of each type. Panel a shows that there is no duration dependence for the transition to P , and we still find that the predicted hazard is greater for agency workers at all durations. For instance, for the 10th week in employment, the hazard is 0.48% for agency workers and 0.18% for direct-hire temps. These figures are 0.50% and 0.20% when unobserved heterogeneity is not accounted for. A similar pattern is found when we look at the exists to accounting for unobserved heterogeneity increases the positive efect of THA employment on the probability of entering unemployment. As to the exit to A, we find that agency workers are much more likely to make this transition than direct-hire temps, although in this case the magnitude of the marginal efect is smaller when unobserved heterogeneity is accounted for. Similarly, direct-temps are again more likely to experience these transitions to D, but the magnitude of this efect is also smaller now.

¹⁸The marginal efects by worker characteristics are available upon request.

Table 3: Heterogeneity Parameters

Notes: The distribution of the unobserved heterogeneity is specified with two mass points for each risk: with probabilities . To ensure that the estimated probabilities are between 0 and 1, we apply a logistic transformation, so that the estimated parameters are log( , with standard errors of 0.422, 0.082, 0.132, 0.085 for j=P,U,D,A.

In terms of the estimated average marginal efects by period, the results are again qualitatively similar in the models with and without unobserved heterogeneity. As Table 10 shows, the marginal efect of THA employment on the exit rate to is again pro-cyclical, and in general the magnitude is very similar. For instance, after 3 months of employment, the average prob ability of entering P is 0.4 pp higher for agency workers than for direct-hire temps during the expansion, while this figure falls to 0.3 pp during the recession and then increases again to 0.5 pp during the recovery. These figures are 0.5, 0.3 and 0.5 pp in the models without unobserved heterogeneity (see Table 2).

As to the transitions to , we still find that the marginal efect of THA employment increased after 2008 and has not decreased during the recovery, though the magnitude of the efects is slightly higher when unobserved heterogeneity is accounted for. For instance, after 3 months of employment, THA employment increases the probability of exiting to U by 0.7 pp before

2008, while the corresponding estimate after 2008 is around 1.1-1.2 pp. In the model without unobserved heterogeneity, these figures are 0.4 and 0.6 pp, respectively. For exits to A and D, we also find a similar pattern by period in the model with and without unobserved heterogeneity. Yet, in contrast with the exit to U , the estimation suggests that the associated marginal efects are smaller (in absolute value) when unobserved heterogeneity is accounted for.

7 Conclusions

Our main findings can be summarized as follows. (1) THA employment has a positive efect on the probability of switching from temporary employment to unemployment, and this probability displays negative duration dependence. Moreover, this efect is stronger during the recession than during the expansion, but it has not decreased during the recovery period. Part of the increase in its magnitude then seems to be structural (rather than cyclical) and may be related with the progressive liberalization of the THA sector, which intensified after the recession, and the fact that THA are mainly placing low-skill workers in low-quality jobs with increasingly higher turnover. (2) Transitions into temporary employment are more likely for agency workers if the new contract is also a THA contract, and for direct-hire temps if it is a regular contract. These diferential efects follow a similar pattern along the business cycle as in the case of the transition to unemployment. (3) THA workers are more likely to transition to permanent employment than direct-hire temps at all durations, the efect being stronger during expansion and recovery periods. While this is consistent with the view that agency jobs have a stepping stone efect when compared with regular temporary jobs—even for low-qualified workers—these transitions are very infrequent for both. (4) Accounting for individual unobserved heterogeneity in the above hazard rates changes the magnitude of the marginal efects, but it does not alter our conclusions.

The results in this paper can be related to other estimates reported in the literature for Spain. Our estimates difer from those of Amuedo-Dorantes et al. (2008), who use a propensity score matching approach to evaluate the impact of agency work on the temporary workers’ posterior likelihood of being hired on a permanent basis. They find that workers who have had a THA job at some point during their sample period (1998-2004) have a lower probability of being hired on a permanent basis after their temporary assignment than their direct-hire counterparts. Their approach is similar to ours in the sense that they also compare temporary workers, with and without a THA contract. However, they account neither for duration dependence, which seems crucial given the diferences between the two worker groups, nor for the efect on other destination states. García-Pérez and Muñoz-Bullón (2005a) estimate duration models for the period 1990-99, but they analyze transitions out of unemployment and out of employment. They find that having previous experience in THA employment has a positive impact on the likelihood of leaving unemployment, but it also implies shorter job tenure while in employment. The main diference with our approach is that we look explicitly at the efect of current THA employment on the transition out of temporary employment and thus focus on temporary workers only. We also study the efect for diferent destination states, and for a longer period which covers the expansion, the recession and the recovery. In addition, we select a highly homogeneous group of low qualified workers and study shorter durations.

Our analysis reveals the importance of accounting for short duration dependence and for diferent risks when analyzing the efect of THA contracts on the hazard of exiting temporary employment. We identify mainly two types of efects when we consider transitions to alternative employment states: the aforementioned small stepping stone efect on the hazard of entering permanent employment, but also a positive efect on the probability associated to a repeated spell in THA employment. The former efect is consistent with the argument that THA may act as a screening device that improves the assignment of workers to jobs. The latter efect can be explained by two channels. The first is a direct channel, as agency employment is associated with a large risk of transiting to another agency job. The second is an indirect channel, as agency workers face also higher unemployment risk, which can lead them to re-start the job search through a THA in order to exit unemployment. This higher unemployment risk might be related to diferences in the kind of jobs employers seek to fill when they hire workers directly and when they do it through a THA, in particular, if the latter are on average more unstable jobs characterized by higher separation rates. In a companion paper (Carrasco et al., 2022), we propose a theoretical mechanism (along these lines) that is consistent with some of our key empirical findings. Unfortunately, our dataset does not provide information about the jobs that agency workers perform at client firms. Yet we do have indirect (descriptive) evidence on wages. Conditional on duration, THA jobs in our sample tend to pay higher entry wages. This is consistent with the idea that agency workers ought to be compensated for the higher turnover risk they face. A full empirical analysis focusing on wages is outside the scope of this paper, and is left for future work.

References

1. Amuedo-Dorantes, C., M. Malo, and F. Muñoz-Bullón (2008). The Role of Temporary Help Agency Employment on Temp-to-Perm Transitions. Journal of Labor Research 29 (2), 138– 161.
2. Antoni, M. and E. J. Jahn (2009). Do Changes in Regulation Afect Employment Duration in Temporary Work Agencies? ILR Review 62 (2), 226–251.
3. Autor, D. H. (2009). Studies of Labor Market Intermediation. NBER Books. National Bureau of Economic Research, Inc.
4. Autor, D. H. and S. Houseman (2005). Do Temporary Help Jobs Improve Labor Market Outcomes for Low-Skilled Workers? Evidence from ’Work First’. Technical report.
5. Bentolila, S., P. Cahuc, J. J. Dolado, and T. Le Barbanchon (2012). Two-tier Labour Markets in the Great Recession: France Versus Spain. The Economic Journal 122 (562), F155–F187.
6. Bover, O., M. Arellano, and S. Bentolila (2002). Unemployment Duration, Benefit Duration and the Business Cycle. The Economic Journal 112 (479), 223–265.
7. Bover, O. and R. Gómez (2004). Another Look at Unemployment Duration: Exit to a Permanent vs. a Temporary Job. Investigaciones Económicas 28 (2), 285–314.
8. Bullón, F. M. and E. C. Rodes (2004). Temporalidad y señalización en el mercado de trabajo: El papel de las empresas de trabajo temporal. Cuadernos de Economía y Dirección de la Empresa (18), 35–67.
9. Carrasco, R., I. Gálvez-Iniesta, and B. Jerez (2022). A Theory of Agency Work. Mimeo, UC3M.
10. Carrasco, R. and J. I. García-Pérez (2015). Employment Dynamics of Immigrants Versus Natives: Evidence from the Boom-Bust Period in Spain, 2000–2011. Economic Inquiry 53 (2), 1038–1060.
11. Countouris, N., S. Deakin, M. Freedland, A. Koukiadaki, and J. Prassl (2016). Report on Temporary Employment Agencies and Temporary Agency Work. Technical report, International Labor Ofice: Geneva.
12. Dolado, J. J., C. García-Serrano, and J. F. Jimeno (2002). Drawing Lessons from the Boom of Temporary Jobs in Spain. The Economic Journal 112 (480), F270–F295.
13. Felgueroso, F., J. García-Pérez, M. Jansen, and D. Troncoso-Ponce (2018). The Surge in Short-Duration Contracts in Spain. De Economist 166 (4), 503–534.
14. Felgueroso, F., J. I. García-Pérez, M. Jansen, and D. Troncoso Ponce (2017). Recent Trends in the Use of Temporary Contracts in Spain. FEDEA Working Paper 11743, FEDEA.
15. Fudge, J. and K. Strauss (2013). Temporary Work, Agencies and Unfree Labour: Insecurity in the New World of Work. Routledge, London.
16. García-Pérez, J. I., I. Marinescu, and J. Vall Castello (2019). Can Fixed-Term Contracts Put Low Skilled Youth on a Better Career Path? Evidence from Spain. The Economic Journal 129 (620), 1693–1730.
17. García-Pérez, J. I. and F. Muñoz-Bullón (2005a). Are Temporary Help Agencies Changing Mobility Patterns in the Spanish Labour Market? Spanish Economic Review 7 (1), 43–65.
18. García-Pérez, J. I. and F. Muñoz-Bullón (2005b). Temporary Help Agencies and Occupational Mobility. Oxford Bulletin of Economics and Statistics 67 (2), 163–180.
19. Harrington, J. W. and N. Velluzzi (2008). Labour market intermediation. In Globalising Worlds and New Economic Configurations. Routledge.
20. Heckman, J. and B. Singer (1984). A Method for Minimizing the Impact of Distributional Assumptions in Econometric Models for Duration Data. Econometrica, 271–320.
21. Heinrich, C. J., P. R. Mueser, and K. R. Troske (2005). Welfare to Temporary Work: Implications for Labor Market Outcomes. The Review of Economics and Statistics 87 (1), 154–173.
22. Ichino, A., F. Mealli, and T. Nannicini (2005). Temporary Work Agencies in Italy: A Springboard Toward Permanent Employment? Giornale degli Economisti 64 (1), 1–27.
23. Jahn, E. J. and M. Rosholm (2014). Looking Beyond the Bridge: How Temporary Agency Employment Afects Labor Market Outcomes. European Economic Review, 108–125.
24. Jenkins, S. P. (1995). Easy Estimation Methods for Discrete-Time Duration Models. Oxford Bulletin of Economics and Statistics 57 (1), 129–138.
25. Kvasnicka, M. (2009, April). Does Temporary Help Work Provide a Stepping Stone to Regular Employment? In Studies of Labor Market Intermediation, NBER Chapters, pp. 335–372. National Bureau of Economic Research, Inc.
26. Malo, M. A. and F. Muñoz-Bullón (2008). Temporary Help Agencies and Participation Histories in the Labour Market: a Sequences-Oriented Approach. Estadística Española 50, 25–65.
27. Narendranathan, W. and M. B. Stewart (1993). How Does the Benefit Efect Vary as Unem ployment Spells Lengthen? Journal of Applied Econometrics 8 (4), 361–381.
28. Theodore, N. and J. Peck (2002). The Temporary Stafing Industry: Growth Imperatives and Limits to Contingency. Economic Geography 78 (4), 463–493.
29. Van den Berg, G. J. (2001). Duration Models: Specification, Identification and Multiple Dura tions. In Handbook of Econometrics, Volume 5, pp. 3381–3460. Elsevier.
30. Voss, E., K. Vitols, N. Farvaque, A. Broughton, F. Behling, F. Dota, S. Leonardi, and F. Naedenoen (2013). The Role of Temporary Agency Work and Labour Market Transitions in Europe: Institutional Frameworks, Empirical Evidence, Good Practice and the Impact of Social Dialogue. Technical report, Hamburg: Eurociett/UNI Europa Project.
31. World Employment Confederation (2021). Economic Report 2021. Technical report.
32. Zijl, M., G. van den Berg, and A. Heyma (2011). Stepping Stones for the Unemployed: the Efect of Temporary Jobs on the Duration until (Regular) Work. Journal of Population Economics 24 (1), 107–139.

Appendices

A Other Tables and Figures

Table 4: Composition of workers by contract type

Notes: Sample of salaried workers aged 16 to 65. We pool all contracts signed in 2017. Source: MCVL.

Table 5: Contract shares by worker/job characteristics

Notes: Sample of salaried workers aged 16 to 65. We pool all contracts signed in 2017. Source: MCVL.

Figure 7: Sectoral composition of agency work (2017)

THA workers over all employed

THA contracts over new contracts

Notes: The vertical line in the right panel figure stands for the average share of THA contracts across all sectors. Own elaboration using data from the THA Statistics, Spanish Social Security.

Table 6: Variables Definition

Table 7: Occupation Classification

Note: The social security asks employers to assign workers into one of ten occupation categories, which are closely related to the level of formal education required for the job. We divide them into five broad occupation groups

Table 8: Estimated coeficients of the single risk and competing risk models

Note: Standard errors in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.010

Table 9: Estimated coeficients, CR models with unobserved heterogeneity

Note: Standard errors in parentheses. ^∗ p < 0.10, ^∗∗ p < 0.05, ^∗∗∗ p < 0.010

Figure 8: Marginal efect of THA by period Panel a: To a Permanent job

Panel b: To unemployment

Panel c: To direct-hire temp job Panel d: To THA job

Notes: The figures display the average marginal efect at each duration, by period. Apart from controlling by agency employment and the interaction between agency employment and duration, all models include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, in direct-hire temporary employment and in agency employment and year fixed efects. The dashes lines correspond to 95 percent confidence intervals. Source: MCVL, 2005-2017.

Figure 9: Predicted hazard rate from the single risk model, THA vs direct temps

Notes: The figure plots the average predicted probability of agency employment at each duration. Apart from controlling by type of contract and the interaction between agency employment and duration, the model includes controls for age, nationality, part-time work, skills, experience in permanent, agency and direct-hire temporary employment, and year fixed efects. The dashes lines correspond to 95 percent confidence intervals. Source: MCVL, 2005-2017.

Table 10: Marginal efect of THA by period, model with unobserved heterogeneity

Notes: The table display the average marginal efects of agency employment by period for a selected set of job duration. Letters E, C, and R stands for the expansion (2004-2007), crisis (2008-2012) and recov ery periods (2013-2017), respectively. For durations that are aggregated, we compute the average. Apart from controlling by agency employment and the interaction between agency employment and duration, all models include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, in direct-hire temporary employment and in agency employment and year fixed efects. Source: MCVL, 2005-2017.

Figure 10: Predicted hazard rates of the model with unobserved heterogeneity Panel a: To a Permanent job Panel c: To direct-hire temp

Panel b: To unemployment Panel d: To THA job

Notes: The figures display the average marginal efects of agency employment at each duration. Apart from controlling by agency employment and the interaction between agency employment and duration, all models include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, in direct-hire temporary employment and in agency employment and year fixed efects. Source: MCVL, 2005-2017.

Figure 11: Marginal efect of THA by period, model with unobserved heterogeneity Panel a: To a Permanent job

Panel b: To unemployment Panel d: To THA job

Panel c: To direct-hire temp

Notes: The figures display the average marginal efects of agency employment at each duration, by period. Apart from controlling by agency employment and the interaction between agency employment and duration, all models include controls for age, nationality, part-time, skills, type of contract, experience in permanent employment, in direct-hire temporary employment and in agency employment and year fixed efects. Source: MCVL, 2005-2017.